Convolution Neural Network-Based Sensitive Security Parameter Identification and Analysis

Abstract

When the logic of a pseudo-random number generator is followed by standards, the security of the generator depends on the entropy sources used in the seed construction, which are constructed through noise sources; thus, the noise sources are extremely important in cryptographic applications. The operation process of a cryptographic random number generator can be divided into an entropy collection stage and a pseudo-random number generation stage. To ensure the security, the noise sources used to construct the entropy source must be securely collected. If a cryptographic analyst physically acquires a cryptographic module in the form of a black box, it may be possible to predict the cryptographic random number to be used in the future by analyzing the random number generation process of the module. Thus, even if the cryptographic algorithms and protocols are securely designed and implemented without vulnerabilities, if the random number generation process is analyzed, the cryptographic system may be exposed. Noise sources can be identified through information generated when they are collected. If the noise sources can be identified during the stage of collecting sources that provide unpredictability to the cryptographic module during random number generation, future values may be predicted according to the data characteristics. We identify noise sources that are used as entropy sources with our convolution neural network model. Therefore, we establish attack scenarios in which random numbers can be analyzed by identifying data used in the model learned through this study when a random cryptographic module is obtained.